Data-driven Monitoring and Quality Prediction for Nonlinear Batch/Multi-grade Productive Processes

Author:Liu Jing Xiang

Supervisor:liu tao


Degree Year:2019





With the rapid development of modern industrial technologies and the diversity of market demands,various chemical processes become more and more complex.Multi-phase batch processes and multi-grade processes are increasingly built up to realize better product quality and economic benefit,therefore bringing higher requirements and challenges to online monitoring,fault detection and quality prediction.With the rapid development and large engineering applications of computer and sensor technologies,data-driven process analysis technologies and modeling methods become effective approaches to solve the above problems,which have been quickly developed in the past decade.In particular,multivariate statistical modeling methods have been widely used for process monitoring and quality prediction.However,there are a number of idealized assumptions and constraints in the existing data-driven modeling and process monitoring methods,such as single process mode,normal distribution of sampled data,linear relationship between variables etc,which may be not suitable or even unable to be applied to multi-phase batch processes and multi-grade processes.This dissertation focuses on multi-phase partition and fault detection for nonlinear batch processes,feature extraction,fault detection and quality prediction for multi-grade processes,quality prediction and operation optimization for cooling crystallization processes.The main contents include;(1)Concerning the multi-phase problem associated with nonlinear batch processes,a moving window based multi-phase partition method is proposed to improve monitoring performance for these batch processes.Firstly,the original process data are mapped into a high-dimensional feature space by using kernel functions,where phase partition is conducted sequentially in order to maintain the time sequence of batch process data.Subsequently,phase partition algorithms are developed for even-and uneven-length batch processes,respectively.Based on the partition results,the corresponding monitoring models are established for each phase,so as to improve online monitoring accuracy and reliability.The proposed method is applied for phase partition of a numerical example and an injection molding batch process,respectively,along with fault detection for a penicillin fermentation process.The results manifest the effictiveness and advantages of the proposed method by comparison with other methods given in the recent references.(2)To address the problem of transition phases existing between individual phases of multi-phase batch processes,a sequential phase partition method is proposed based on Gaussian mixture models.An independent Gaussian model is adopted to describe a stable phase,while a transition phase is described by a mixture of the Gaussian models of two adjacent stable phases.Meanwhile,phase partition is conducted based on local sequential process data,such that the computation cost could be significantly reduced while maintaining the time sequence of partitioned phases.The corresponding probability models for each stable or transition phase are established for improving online monitoring performance.The phase partition and fault detection performance of the proposed method are demonstrated by a numerical example and a penicillin fermentation process,respectively,well indicating its effectiveness and advantages.(3)Concerning the common features and differences between multi-grade processes,together with the problem of insufficient data measured from each process grade arising from frequent shifts between different processes of the same production line,a common feature extraction method is proposed based on the sampled data of different process grades.Then a principal component analysis(PCA)method is used to extract the special feature of each grade.Therefore,each grade of these processes is divided into common part,special part,and residual part.The corresponding monitoring models are established for fault detection,respectively.A numerical example and an industrial polyethylene process are used to demonstrate the effictiveness and advantages of the proposed method.(4)To cope with the difficulty of on-line quality prediction for multi-grade processes,a modeling method is proposed for quality prediction by combining a just-in-time learning strategy and a common feature extraction approach.A just-in-time learning strategy is adopted to handle the nonlinear problem associated with practical multi-grade processes.A novel common feature extraction algorithm is given to determine the common features shared by different process grades,according to the correlation between process variables and quality variables.Then a partial least-squares(PLS)modeling algorithm is used to extract the special features of each grade,respectively.Hence,product quality prediction is conducted by integrating the common and special features of each grade,so as to improve the accuracy and reliability.The effictiveness and advantages of the proposed method are demonstrated by a numerical example and an industrial polyethylene process.(5)For nonlinear batch cooling crystallization processes,a data-driven analysis method is proposed for predicting one-dimensional product crystal size distribution(CSD)or chord length distribution(CLD),based on double-layer basis functions.Two classes of basis functions are adopted,one is the wavelet basis function for reshaping the CSD curve,and the other is the polynomial basis function for weighting the selected wavelet basis functions.An active learning strategy is given to determine the number of wavelet basis functions,while the output prediction error is used to determine the number of polynomial basis functions.By introducing an objective function combining the entropy of the product CSD with the deviation between the desired and predicted CSD for optimization,an optimization method is established for the temperature operation of cooling crystallization.A simulated hen-egg-white lysozyme crystallization process and a real L-glutamic acid cooling crystallization process are used to demonstrate the effectiveness and advantages of the proposed method for optimizing the temperature operation.